The operation of a complex process facility requires the assimilation of a large amount of data generated frequently from sensors, inferring the state of the process from these data and executing regulatory and supervisory actions to ensure safety and efficiency. While much of the operation of the modern complex process facility is automatically controlled, the execution of the process during transitions requires extensive human intervention and supervision. These operator interventions during transitions include starting or stopping process units such as pumps and compressors, disabling or enabling controllers, and reconfiguring units and controllers with different parameter settings. Transitions result in substantial changes in the values of one or more process variables. During transitions, some of these plant variables may therefore take values substantially different and outside their normal, steady state range. Because of the large changes and the unusual values of the process variables, plant computer programs that are configured for only one operating condition generate incorrect results. One example of this is the alarm system that would generate spurious alarms if it were not specifically configured for the transition. Other such computer programs include those for advanced control, unit optimization, etc. Operators therefore disable and/or ignore the results of such plant computer programs during transitions. The operator actions, mentioned above, that bring about a transition have to be performed at a specified time dependent on the status and conditions of the plant. Failure in this can lead to abnormal situations.
D. H. Hwang and C. Han, “Real-time Monitoring For A Process With Multiple Operating Modes”, Control Engineering Practice, 7, 891-901, 1999, proposed a statistical method for monitoring processes with multiple operating modes. Their technique is applicable only for operation modes a that share common characteristics and do not introduce a significant amount of nonlinearities in the process behavior. Many common transitions such as a cold startup of a reactor or distillation column, decoking of a furnace, swinging of units for product or feed grade changes, and others result in significant nonlinear changes and cannot be monitored by their approach.
Process transitions are typically carried out by operators by following a standard operating procedure. A few techniques exist for monitoring the execution of the operating procedures in a process plant.
A method suggested in U.S. Pat. No. 5,511,004, issued April, 1996 to Dubost et al., establishes a reference state and a current state for an industrial evolutionary process from physical parameters measured on all the equipment items employing the evaluating process. These two states are compared, parameter by parameter, by resorting to fuzzy logic for classifying the quantities, and a diagnosis is established using expert rules.
U.S. Pat. No. 5,070,468, issued December, 1991 to Niinomi et al., presents a fault diagnosis system. The normal range of process data is stored in the system and is employed for online comparison to determine if the process is normal or not. Patterns exhibited by the process variables for different fault patterns are detected and recorded in the system. When a fault is identified, the pattern is compared with this stored database and the fault is diagnosed.
U.S. Pat. No. 5,392,320, issued February, 1995 to Chao et al., provides a monitoring system for a core of the nuclear reactor. Some process variables that need to be continuously monitored are identified. A database is maintained for normal and abnormal operation. Comparison of online data is done with this database to identify the plant situation.
An approach to monitoring a transient is described in U.S. Pat. No. 4,678,622, issued July, 1987 to Rowe et al. This makes use of the known phenomenon that during the startup in a nuclear reactor, the neutron density increases. However, there could also be an increase in the neutron density due to some abnormality. A system is built to detect the abnormality. If the startup is done normally, the neutron density increases slowly, whereas if there is a fault, there is an exponential increase. This difference is used for identifying the fault during startup.
A method and apparatus for fault diagnosis is described in U.S. Pat. No. 5,099,436, issued March, 1992 to McCown et al. In this system, the online values of variables are continuously monitored. If there is any deviation from normal, it is mapped to any pre-trained event (fault). It also consists of a symptom-fault model that can determine the cause of the failure.
U.S. Pat. No. 5,023,045, issued June, 1991 to Watanabe et al., describes a neural network based technique for detecting malfunctions in plants. The neural network is trained with normal and different runs of data with malfunctions. So this system, when used with online run, can identify trained malfunctions. An application on a nuclear power station has been discussed. Since this application requires training, it can function only for known faults for which the network has been trained.
A few other techniques also exist for monitoring the operating procedures for nuclear and thermal power plants. However, all of these techniques employ raw sensor values in their analysis.
There exist some techniques in published literature which use signed directed graphs (see Iri, M., Aoki, K., O'Shima, E. and Matsuyama, H., “An Algorithm For Diagnosis Of System Failures In Chemical Processes”, Computers and Chemical Engineering, 3, 489-493, 1979) and trends (see Rengaswamy, R. and Venkatasubramanian, V., “A Syntactic Pattern Recognition Approach For Process Monitoring And Fault Diagnosis”, Engineering Applications Of Artificial Intelligence, 8(1), 35-51, 1995; Vedam, H. and Venkatasubramanian, V., “A Wavelet Theory Based Adaptive Trend Analysis System For Process Fault Diagnosis”, Proceedings Of The American Control Conference, 309-313, 1997) for monitoring steady state processes. During process transitions, the process is in a dynamic state and variable profiles change significantly. For monitoring such process behavior, none of the existing techniques (such as neural networks, signed directed graph or trend-based methods) could be used.
First, consider a signed digraph based technique. During process transitions, since interactions between the different variables vary with process operating conditions and time, a general cause-effect relation between process variables could not be obtained. So signed digraph based techniques cannot be used for monitoring process transitions.
In the trend-based approaches, the evolution of a process variable is classified based on its shape into slowly increasing, drastically increasing, constant, etc., by calculating the second order derivatives of the process variables. These are termed as second-order trends. First order derivatives of process variables result in trends which are classified into increasing, decreasing and constant. These are termed as first-order trends. Neither the second-order nor the first-order trend-based approaches can be applied to monitor process transitions as described below. As an example, consider the transition “startup of a reactor” that is performed in three main phases: (1) charging of reactants at room temperature (30° C.) into the reactor, (2) ramping up the temperature from 30° C. to 75° C. over 1.5 hours, and (3) maintaining the temperature at 75° C. Each process variable could display different trends during different phases. The temperature sensor in the above example would display a constant trend in phase 1, an increasing trend in phase 2, and a constant trend again in phase 3. Existing trend-based approaches are designed for fault detection in steady state operation and map a fixed trend (such as increasing temperature) occurring at any time with a fixed process state (such as a runaway reaction). This can lead to wrong results when used to monitor process transitions. In the above example, while the increasing trend in temperature would be diagnosed correctly as due to a runaway reaction if the process is in phase 3, an increasing trend in the same temperature variable during phase 2 would be misdiagnosed as due to a runaway reaction. Due to this and other limitations discussed later, existing trend based techniques can not be applied for monitoring transitions.